How AI should read a customer email — and what it must refuse to do

A customer writes: *"Hi — need 500 business cards like the last ones, both sides, matt, by Friday."*

Any language model can turn that into a quote form. Quantity 500. Sides 2. Finish matt lamination. Due Friday. It will look complete, it will look confident, and in a print shop it is a reprint waiting to happen.

What the model actually knows

Look again at what that email contains.

It states the quantity, the sides and the finish. It does not state the size, the stock, the weight or the colours. And it gestures at all of them with three words — "like the last ones" — which is a reference, not a specification.

A model that fills those fields anyway is guessing. It might guess well: 85×55 on 350gsm is a reasonable default in most shops. But "reasonable default" is doing a lot of work in a business where the difference between 300 and 350gsm is a customer complaint, and where the last job might have been 90×50 because that customer's designer likes it that way.

Three buckets, not one

The useful design separates every field into one of three states, and shows which is which.

Stated. The customer said it in this email. Quantity 500, two sides, matt. These you can act on.

From history. The system found it by looking at what this customer ordered before. Size 85×55, 350gsm silk, from quote QT-2026-0184 in March. This is a *suggestion*, and it should arrive with its source attached so you can see where it came from and decide whether it still holds. "Like the last ones" is a strong hint — but only the person reading it knows whether the customer means the last ones or the last ones before the rebrand.

Missing. Nothing in the email and nothing in the history. These stay blank and stay visible.

The failure mode of most AI features is collapsing all three into one. Everything arrives filled in, nothing indicates which values were read and which were invented, and the estimator's job silently changes from *entering specs* to *auditing a machine*. That is worse than an empty form, because an empty form is honest about what it does not know.

Two rules that are easy to get wrong

"Same as last time" is a question, not an answer. It is a reference to a previous job, so the fields it implies belong in the missing bucket with the reference noted, not in the confirmed bucket with values copied across. The customer may well mean exactly that. They may also have forgotten they changed the stock in June. Only a human who knows the account can tell.

"No lamination" is not a missing value. It is an explicit decision, and a system that treats it as a gap will nag the estimator to fill in a field the customer has already answered. Absence of a value and a stated absence are different things, and conflating them makes the tool feel stupid in a way people remember.

Then write the email back

The genuinely useful output of the missing bucket is not a red asterisk. It is the reply.

If the system knows exactly which four fields it could not determine, it can draft the message that asks for them — in the shop's own words, naming the specific gaps, ready to send. That turns the weakest part of the process, the two days lost waiting for someone to remember to chase the customer, into a single click on the day the request arrived.

Why print is different

Most industries can absorb a bad guess. A wrong assumption in a CRM produces an awkward sales call. A wrong assumption in a print quote produces 500 cards on the wrong stock, and the cost of that is not the estimate — it is the paper, the plates, the press time and the customer.

So the honest promise for AI in this trade is not that it knows. It is that it shows you what it read, what it inferred, and what nobody has answered yet — and then gets out of the way.

Where we sit

PressCal does the above: it separates stated from historical from missing, shows the source quote and date behind every suggestion, refuses to fill blanks from a reference like "same as last time", and drafts the email asking for what is left. I am telling you because I built it, and because the principle matters more than the product — if you are evaluating anything with AI in it, ask to see the missing list. If there isn't one, the guesses are in there somewhere.